Dynamical Systems and Ergodic Theory

Transcription

1 MATH MATHM606 Dynamical Systems and Ergodic Theory Teaching Block, 07/8 Lecturer: Prof. Alexander Gorodnik PART I: LECTURES -7 course web site: people.maths.bris.ac.uk/ mazag/ds7/ Copyright c University of Bristol 00 & 04. This material is copyright of the University. It is provided exclusively for educational purposes at the University and is to be downloaded or copied for your private study only.

2 Chapter Examples of dynamical systems. Introduction Dynamical systems is an exciting and very active field in pure and applied mathematics, that involves tools and techniques from many areas such as analyses, geometry and number theory and has applications in many fields as physics, astronomy, biology, meterology, economics. The adjective dynamical refers to the fact that the systems we are interested in is evolving in time. In applied dynamics the systems studied could be for example a box containing molecules of gas in physics, a species population in biology, the financial market in economics, the wind currents in metereology. In pure mathematics, a dynamical system can be obtained by iterating a function or letting evolve in time the solution of equation. Discrete dynamical systems are systems for which the time evolves in discrete units. For example, we could record the number of individuals of a population every year and analyze the growth year by year. The time is parametrized by a discrete variable n which assumes integer values: we will denote natural numbers by N and integer numbers by Z. In a continuous dynamical system the time variable changes continuously and it is given a real number t. We will denote real numbers by R. Our main examples of discrete dynamical systems are obtained by iterating a map. Let X be a space. For example, X could be the unit interval [0, ], the unit square [0, ] [0, ], a circle (but also the surface of a doughnut or a Cantor set). Let f : X X be a map. We can think as f as the map which gives the time evolution of the points of X. If x X, consider the iterates x, f(x), f(f(x)),.... Notation... For n > 0 we denote by f n (x) the n th iterate of f at x, i.e. f f f, n times. In particular, f = f and by convention f 0 : is the identity map, which will be denoted by Id (Id(x) = x for all x X). We can think of f n (x) as the status of the point x at time n. We call forward orbit the evolution of a point x. Definition... We denote by O + f (x) the forward orbit of a point x X under iterates of the map f, i.e. O + f (x) := {x, f(x), f (x),..., f n (x),... } = {f n (x), n N}. Here N is the set of natural numbers including 0. This gives an example of a discrete dynamical system parametrized by n N. Do not confuse this notation with the n th derivative, which will be denoted by f (n), or by the n th power, which will not be used! The name orbit cames from astronomy. The first dynamical system studied were indeed the solar system, where trajectory of a point (in this case a planet or a star) is an orbit.

3 Example... Let X = [0, ] be the unit interval. Let f : X X be the map f(x) = 4x( x). For example ( ) { O + f = 3 3, 4 ( ) = , 4 8 ( 8 ) = 3 } 9 9 8,.... Example... Let X be a circle of radius. An example of map f : X X is the (clockwise) rotation by an angle πα, which maps each point on the circle to the point obtained by rotating clockwise by an angle πα. If f is invertible, we have a well defined inverse f : X X and we can also consider backwards iterates f (x), f (x),.... Notation... If f is invertible and n < 0, we denote by f n (x) the n th iterate of f at x, i.e. f f f, n times. Note that even if f is not invertible, we will often write f (A) where A X to denote the set of premimges of A, i.e. the set of x X such that f(x) A. Definition... If f is invertible, we denote by O f (x) the (full) orbit of a point x X under forward and backward iterates of f, i.e. O f (x) := {..., f k (x),..., f (x), f (x), x, f(x), f (x),..., f k (x),... } = {f k (x), k Z}. In this case, we have an example of discrete dynamical system in which we are interested in both past and future and the time is indexed by Z. Even if the rule of evolution is deterministic, the long term behavior of the system is often chaotic. For example, even if two points x, y are very close, there exists a large n such that f n (x) and f n (y) are far apart. This property (which we will express formally later) is known as sensistive dependence of initial conditions. There are various mathematical definitions of chaos, but they all include sensitive dependence of initial conditions. Different branches of dynamical systems, in particular topological dynamics and ergodic theory, provide tools to quantify how chaotic is a systems and to predict the asymptotic behaviour. We will see that often even if one cannot predict the behaviour of each single orbit (since even if deterministic it is too complicated), one can predict the average behaviour. The main objective in dynamical systems is to understand the behaviour of all (or almost all) the orbits. Orbits can be fairly complicated even if the map is quite simple. A first basic question is whether orbits are finite or infinite. Even if the index run through a infinite set (as N or Z) it could happen that O f (x) is finite, for example if the points in the orbit repeat each other. This is the simplest type of orbit. Definition..3. A point x X is periodic if there exists n N\{0}, such that f n (x) = x. If n =, so that we have f(x) = x, we say that x is a fixed point. More in general, if f n (x) = x we say that x is periodic of periodic n or that n is a period for x. In particular, f n+j (x) = f j (x) for all j 0. Example..3. In example.., the point x = 3/4 is a fixed point, since f(3/4) = 4 3/4( 3/4) = 3/4. Example..4. In example.., set α = /4, i.e. we consider the rotation by π/, all points are periodic with period 4 and all orbits consist of four points: the initial points are the points obtained rotating it by π/, π and 3π/. Definition..4. If x is a periodic point, the minimal period of x is the minimum integer n such that f n (x) = x. In particular, if n is the minimal period of x, the points f(x),..., f n (x) are all different than x. Be aware that in some textbook the period of a periodic point x means the minimal period. Definition..5. A point x X is preperiodic if there exists k N, n N \ {0} such that f n+k (x) = f k (x). In this case f n+j (f k (x)) = f j (f k (x)) for all j N. Exercise... Show that if f is invertible every preperiodic point is periodic. Examples of questions that are investigated in dynamical systems are: c University of Bristol 00 & 04

4 Q Are there fixed points? Are there periodic points? Q Are periodic points dense? Q3 Is there an orbit which is dense, i.e. an orbit which gets arbitrarily close to any other point in X? Q4 Are all orbits dense? We will answer these questions for the first examples in the next lectures. More in general, these properties are studied in topological dynamics 3, see Chapter. If an orbit is dense, it visits every part of the space. A further natural question is how much time it spends in each part of the space. For example, let A X be a subset of the space. We can count the number of visits of a segment {x, f(x),..., f n (x)} of the orbit O + f (x) to the set A. If χ A denotes the characteristic function of A, i.e. { if x A χ A (x) = 0 if x / A, let us count the number of visits n Card{0 k < n, such that f k (x) X} = χ A (f k (x)) and divide by n to get the frequency of visits in time n: k=0 n χ A (f k (x)). (.) n k=0 Intuitively, orbit O + f (x) is equidistributed if the frequency in (.) is getting closer and closer, as n increases, to the volume of A (or the length, or the area,... ) 4. This means that the orbit asymptotically spends in each part of the space a time proportional to the volume. Q Are orbits equidistributed? This last question is a main quesition in ergodic theory 5 A priory, not even the existence of a limit of the frequency (.) is guaranteed. One of the main theorems that we will see in Chapter 4, the Birkhoff ergodic theorem, will show that for almost all points the limit exists and guarantee that if the system is enough chaotic (more precisely, ergodic), than the frequency converge to the expected limit. As we will see, questions related to equidistributions have many connections and applications in number theory. Course Outline The course will provide an introduction to subject of dynamical systems, from a pure-mathematical point of view. The first part of the course will be driven by examples. We will present many fundamental examples of dynamical systems, such as rotations of the circle; the doubling map; the baker map; automorphisms of the torus; the Gauss map and continued fractions. 3 Topological Dynamics is a branch of dynamics that investigate the properties of continous maps. 4 More in general, we will have a measure on X (length, area and volume are all examples of measures) which is preserved by the map f and we will ask if the frequency tends to the measure of A. See Chapter 4. 5 Ergodic Theory is a branch of dynamics which investigate the chaotic properties of maps which preserves a measure. 3 c University of Bristol 00 & 04

5 Driven by the examples, we will introduce some of the phenomena and main concepts which one is interested in studying. In the second part of the course, we will formalize these concepts. We will then develop the mathematical background and cover the basic definitions and some fundamental theorems and results in three areas of dynamical systems: Topological Dynamics (Chapter ) Symbolic Dynamics (Chapter 3) Ergodic Theory (Chapter 4) During the course we will also mention some applications both to other areas of mathematics, such as number theory, and to problems as data storage and Internet search engines. We will see that the algorithm for the widely used search engine google is based on an idea which uses the theory of dynamical systems. References We will follow the first four Chapters of []. Many of the arguments are also covered in more depth in [] (which is very elementary and assumes no background) and in [3], which is more abstract and recommended for a more theoretical approach. [ ] M. Brin, G. Stuck Introduction to Dynamical Systems, Cambdrige University Press. [ ] B. Hasselblatt, A. Katok A First Course in Dynamics, Cambdrige University Press. [3 ] M. Pollicott and M. Yuri Dynamical Systems and Ergodic Theory, London Mathematical Society.. Rotations of the circle: periodic points and dense orbits Consider a circle of unit radius. More precisely, we will denote by S the set S = {(x, y) x + y = } R. Identifying R with the complex plane C, we can also write S = {z C z = } = {e πiθ, 0 θ < } C. Consider a counterclockwise rotation R α of angle πα > 0 on the circle (see Figure..). It is given by R α (e πiθ ) = e πi(θ+α) = e πiα e πiθ. We will refer to this formula for R α as multiplicative notation, since rotating by πα in this complex notation is equivalent to multiplying the complex number e πiθ by e πiα. [We can also consider rotations by negative angles πα < 0, with the convention that they represent clockwise rotations by πα.] There is a natural distance d(z, z ) between points on S, which is induced by the arc length distance. The arc length distance between two points is the length of the shortest arc connecting the two points. We will renormalized the arc length distance by dividing by π. Thus, d will denote the arc length distance divided by π. Since θ, θ [0, ), we have θ θ <. If θ θ the arc length is If θ θ, d(e πiθ, e πiθ ) = arc length distance π = π θ θ π = θ θ. d(e πiθ, e πiθ ) = arc length distance π = π π θ θ π 4 c University of Bristol 00 & 04 = θ θ.

6 f(f(x)) f(x) x Figure.: A rotation of S. The rotation preserves this distance, meaning that d(r α (z ), R α (z )) = d(z, z ), for all z, z S. This is clear from the geometric picture, since both points are rotated by the same angle πα. Thus, the rotation of the circle is an example of an isometry, i.e. a map which preserves a distance. There is another alternative way to describe a circle, that will be often more convenient. Imagine to cut open the circle to obtain an interval. Let I/ denote the unit interval with the endpoints identified: the symbol recalls us that 0 are glued together. Then I/ is equivalent to a circle. More formally, consider R/Z, i.e. the space whose points are equivalence classes x+z of real numbers x up to integers: two reals x, x R are in the same equivalence class iff there exists k Z such that x = x + k. Then R/Z = I/ since the unit interval I = [0, ] contains exactly one representative for each equivalence class with the only exception of 0 and, which belong to the same equivalence class, but are identifyed. The map Ψ : R/Z S given by x Ψ(x) = e πix. (.) establishes a one-to-one correspondence between R/Z and S. The distance d given by arc length divided by π, gives the following distance on R/Z, that we will also denote with the same symbol d: Exercise... Consider the map Ψ : R/Z S given in (.). d(x, y) = min x y + m. (.3) m Z (a) Check that Ψ is well-defined on R/Z and establishes a one-to-one correspondence between R/Z and S. (b) Show that the distance d on R/Z induced by the distance d on S, that is the distance defined by satisfy (.3). d(x, y) = d(ψ(x), Ψ(y)), x, y R/Z. We will write equivalenlty [0, ]/ or R/Z to denote the interval with glued endopoints. The counterclockwise rotation R α, under the identification between S and R/Z given by (.), becomes the map R α : R/Z R/Z given by R α = x + α mod, where mod means that we subtract the integer part (for example 3.4 mod = 0.4), hence taking the representative of the equivalence class x + α + Z which lies in [0, ). We call α the rotation number of R α (remark that the rotation angle is πα). More explicitely, if x, α [0, ) we have { x + α if x + α < R α (x) = x + α if x + α. 5 c University of Bristol 00 & 04

7 We call this additive notation (since here the rotation becomes addition mod ). Rotations of the circle display a very different behaviour according if the rotation number α is rational (α Q) or irrational (α R\Q). Let us give the definition of density for an orbit of a rotation: Definition... The orbit O + R α (z ) is dense if for all z S and for all ϵ > 0 there exists n > 0 such that d(r n α(z ), z ) < ϵ. Theorem.. ((Dichotomy for Rotations)). Let R α : R/Z R/Z be a rotation of the circle. () If α = p/q is rational, with p, q Z, all orbits are periodic of period q; () If α is irrational, for every point z S the orbit O + R α (z ) is dense. In both cases the dynamics of the rotation is quite simple to describe: either all orbits are periodic, or all orbits are dense. In the proof we will use the following elementary principle, which is called Pigeon Hole Principle (also known as Shoebox Principle): Pigeon Hole Principle. Given N boxes and N + objects, there is one box which contains at least two objects. Proof of Theorem... Let us first prove (). Let us use the additive notation for R α. Let α = p/q with p, q Z. Then for each x R/Z R q α(x) = x + q p q mod = x + p mod = x. Thus every point is periodic of period q. This proves (). Let us now prove (). In this case it will be convenient to work on S and use multiplicative notation. Assume that α is irrational. Step : Let us first show that the each orbit consist of infinitely many distinct points, or, in other words, that for each z = e πix S, for all m n, R n α(e πix ) R m α (e πix ). Let us argue by contradiction. If R n α(e πix ) and R m α (e πix ) were equal, e πi(x +mα) = e πi(x +nα), thus π(x + mα) = π(x + nα) + πk for some integer k N, thus simplifying mα = nα + k. But since m n, this shows that α = k/(m n), contradicting the assumption that α is irrational. Step : Let us show that for any ϵ > 0 there exists two distinct points R n α(z ), R m α (z ) in the orbit O + R α such that d(r n α(z ), R m α (z )) N < ϵ. Let N be big enough so that /N < ϵ. Consider the points z, R α (z ),..., R N α (z ). Since as we proved before are all points in any orbit are distinct, these are N + distinct points. Thus, if we divide the cirle into N arcs of equal length, it follows by the Pigeon Hole principle that at least two of the first N + distinct points of the orbit have to fall inside the same arc, that is there exists n, m with 0 n < m N so that d(r n α(z ), R m α (z )) N < ϵ. Step 3: Finally, to show that the orbit of z S is dense, we have to show that for each z S and ϵ > 0 there is a point of O + R α (z ) inside the ball B(z, ϵ). Consider now Rα m n. We claim that it is again a rotation by an angle smaller than πϵ. First of all, we see that R m n α (z ) = e πi(m n)α z = R (m n)α (z ), 6 c University of Bristol 00 & 04

8 so that Rα m n is a rotation by an angle π(m n)α (counterclockwise or clockwise according to the sign). By Step and using that R α and hence also Rα m are isometries, we have that d(rα n m (z ), z ) = d ( Rα m ( R n m α (z ) ), R m (z ) ) = d(rα(z n ), Rα m (z )) N < ϵ. This shows that the rotation Rα m n moves points by an arc of length less than π/n, i.e. by distance less than /N < ϵ. Thus, if we consider the points in the orbit O R m n α, that is z, R (m n) α (z ), Rα (m n) (z ), Rα 3(m n) (z ),..., their spacing on S is less than πϵ, or in other words the distance between two consecutive points is less than ϵ (recall that the distance is the arc length divided by π). Thus, there will be a j > 0 such that Rα j(m n) (z ) enters the ball B(z, ϵ). Since Rα j(m n) (z ) belongs to O Rα (z), this concludes the proof that every orbit is dense. Exercise... Prove that if α = p/q and (p, q) = i.e. p and q are coprime, then q is the minimal period, i.e. for each x R/Z we have R k α(x) x for each k < q. Remark... If α = p/q and (p, q) = then p gives the winding number, i.e. the number of turns that the orbit of any point does around the circle S before closing up. Exercise..3. Let X = R/Z R/Z and let f(x, x ) = (R α (x ), R α (x )). where α, α are real numbers. Assume that at least one of α, α is irrational. (a) Write down the the n th -term of the orbit O + f (x) of x = (x, x ) under f and show that all points in the orbit O + f (x) of x under f are distinct. (b) Show that for any x = (x, x ) R and any N positive integer there exists n N such that d(f n (x), x) N, where d denotes the distance on X induced by the Euclidean distance on R R..3 The Doubling Map: coding and conjugacy Consider the following map on [0, ), known as doubling map: f(x) = x mod = { x if 0 x < /, x if / x. (.4) The graph is shown in Figure.. The map is well defined also on R/Z = [0, ]/. To check that, we have to check that the points 0 and which are identified have equivalent images. But this is true since f() = and f(0) = 0 and hence f(0) = f() mod. So we can think of f as a map on R/Z = I/. Since we saw that R/Z is identified with S (via the correspondence given by x e πix, see.), we can see f in multiplicative coordinates as a map from S S given by f(e πiθ ) = e πiθ = (e πiθ ). (.5) Thus the angles are doubled and this explains the name doubling map. Moreover, one can see that the map f on S is continuous. 7 c University of Bristol 00 & 04

9 x x 0 / Figure.: The graph of the doubling map. Note that f is not invertible: each point has two preimages: { y f (y) =, y + }. Note also that f expands distances. If d(x, y) < /4, then Let us try to answer the following questions for f: Q. Are there periodic points? Q. Are there points with a dense orbit? d(f(x), f(y)) = d(x, y). In order to answer these questions, we will show two powerful techniques in dynamical systems, conjugacy and coding..3. Conjugacy and semi-conjugacy Let X, Y be two spaces and f : X X and g : Y Y be two maps. Definition.3.. A conjugacy between f and g is an invertible map ψ : Y X such that ψ g = f ψ, i.e. for all y Y ψ(g(y)) = f(ψ(y)). Since ψ is invertible, we can also write g = ψ f ψ. The relation ψ g = f ψ is often expressed by saying that the diagram here below commutes, i.e. one can start from a point in y Y in the left top corner and indifferently apply first the arrow g and then the arrow ψ on the right side of the diagram or first the arrow ψ on the left side and then the arrow f and the result is the same point in the bottom right corner. g Y Y ψ ψ f X X In the following we will often drop from the notation for composition, and write gf instead of g f (as there is usually no danger of confusion with the notation of a product of two functions). 8 c University of Bristol 00 & 04

10 Lemma.3.. If f and g are conjugated by ψ, then y is a periodic point of period n for g if and only if x = ψ(y) is a periodic point of period n for f. Exercise.3.. Check by induction that if ψg = fψ, then ψg n = f n ψ. g n = ψ g n ψ. Proof. Assume that g n (y) = y. Then, by the exercise above If ψ is invertible, we also have f n (ψ(y)) = ψ(g n (y)) = ψ(y), so ψ(y) is a periodic point of period n for f. Conversely, assume that f n (ψ(y)) = ψ(y). Then since ψ is invertible, by the previous exercise we also have g n = ψ g n ψ. Thus g n (y) = ψ (f n (ψ(y))) = ψ (ψ(y)) = y, so y is periodic of period n for g. Thus, if the periodic points of the map g are easier to understand than the periodic points of the map f, through the conjugacy one can gain information about periodic points for f. We will see that this is exactly the case for the doubling map. Definition.3.. A semi-conjugacy between f and g is a surjective map ψ : Y X such that ψg = fψ, i.e. for all y Y ψ(g(y)) = f(ψ(y)), or, equivalently, such that the diagram below commutes: g Y Y ψ ψ X f X In this case we say that g : Y Y is an extension of f : X X and that f : X X is a factor of g : Y Y. Exercise.3.. If f and g are semi-conjugated by ψ and y is a periodic point of period n for g, then ψ(y) is a periodic point of period n for f. Remark.3.. There are examples of f and g which are semi-conjugated by ψ and such that ψ(y) is a periodic point of period n for f, but y is not a periodic point for g. We will see such an example using the baker map in a few classes..3. Doubling map: semi-conjugacy and binary expansions We will define a (semi)conjucagy between the doubling map and an abstract space that will help us understand points with periodic and dense orbits. Given x [0, ], we can express x in binary expansion, i.e. we can write x = i= where x i are digits which are either 0 or. Binary expansions are useful to study the doubling map because if we apply the doubling map: f(x) = x mod = i= x i i x i i mod = x 0 + i= x i i mod = and if we now change the name of the index, setting j = i, we proved that if x = i= x i i, then f(x) = i= x i i = j= i= x i i x j+ j (.6) 9 c University of Bristol 00 & 04

11 i.e. the binary expression of f(x) is such that the digits are shifted by. Let us construct a map on the space of digits of binary expansion which mimic this behavior. Let Σ + = {0, } N be the set of all sequences of 0 and : Σ + = {(a i ) i=, a i {0, }}. The points (a i ) i= Σ+ are one-sided sequences of digits 0,, for example a point is 0, 0,,, 0,, 0, 0, 0,,, 0,,,,, 0, 0,, 0,, 0... The shift map σ + is a map σ + : Σ + Σ + which maps a sequence to the shifted sequence: σ((a i ) i=) = (b i ) i=, where b i = a i+. The sequence (b i ) i= is obtained from the sequence (a i) i= by dropping the first digit a and by shifting all the other digits one place to the left. For example, if (a i ) i= = 0, 0,,, 0,, 0, 0, 0,,, 0,,,,, 0, 0,, 0,, 0... (b i ) i= = 0,,, 0,, 0, 0, 0,,, 0,,,,, 0, 0,, 0,, 0... Note that σ + is not invertible, because if we know σ((a i ) i= ) we cannot recover (a i) i= information about the first digit. Define the following map ψ : Σ + [0, ]. For each (a i ) i= Σ+ set since we lost the ψ((a i ) i=) = i= a i [0, ]. i The map is well defined since the series a i i= i i= which is convergent. The map ψ associates to a i sequence of 0 and a number in [0, ] which has a i as digits of the binary expansion. One can see that ψ is surjective since each real x [0, ] has a binary expansion (in the next section,.3.3, we will show how to produce a binary expansion). On the other hand, it is not injective, since there are numbers which have two binary expansions. In the same way that in decimal expansion we can write = , binary expansions which have an infinite tails of yield the same number that an expansion with a tail of 0, for example = + 0 i but also i= = i. i= One can check that these ambiguity happens only for rational numbers of the form k/ n, whose denominator is a power of, that we will call dyadic rationals. For all other numbers ψ is a bijection. Exercise.3.3. Let D be the set of all dyadic rationals, that is all rational numbers of the form { } k D = n, 0 k < n, n N\{0}. Show that any dyadic rational has two binary expansions, one ending with a tail of 0s and the other ending with a tail of s (that is ending in 0, 0, 0,... or,,,... ). Proposition. The map ψ : Σ + R/Z is a semi-conjugacy between the shift map σ + : Σ + Σ + and the doubling map f : R/Z R/Z. Proof. Let us check that ψ is surjective. Since every point x [0, ] has at least one binary expansion, we can write x = i= a i/ i where a i {0, }. Thus, the sequence a = (a i ) i= Σ+ of its binary digits is such that ψ(a) = x by definition of the ψ. 0 c University of Bristol 00 & 04

12 Thus, we are left to prove that the following diagram commutes: Σ + σ + ψ R/Z Σ + ψ f R/Z Take any (a i ) i= Σ+. Let us first compute ψ(σ + ((a i ) i= )): ψ(σ + ((a i ) i=)) = ψ((b i ) i=) = since b i = a i+. Let us now compare with f(ψ((a i ) i= )): ( ) f(ψ((a i ) a i i=)) = f i = i= i= i= b i i = i= a i i mod = a i+ i, in virtue of the computation done in (.6). Thus the results are the same. This concludes the proof. Remark: The semi-conjugacy ψ between the doubling map f and the shift σ + is very close to be an actual conjugacy, since ψ fails to be a conjugacy only on a countable set. If one removes a countable set of points, one gets a conjugacy. Another way to obtain a conjugacy on the full space is explained in Extra A.3.. Exercise.3.4. Let Σ Σ be obtained removing all sequences with a tail of 0s or with a tail of s Verity that ψ : Σ R/Z is injective. On which subset X R/Z we have that ψ : Σ X is also surjective and thus is a one to one correspondence? [Hint: recall the previous exercise on dyadic rationals.] The map ψ : Σ + R/Z defined above sends a sequence in 0 and to the corresponding binary expansion. Conversely, given a point x [0, ], we are going to see how one can construct a binary expansion for x using symbolic coding. Periodic points Given a dynamical system f : X X, we will denote by P er n (f) the set of all periodic points of period n for f and by P er(f) the set of all periodic points, that is P er n = {x X, f n (x) = x}, P er(f) = j= a j+ j P er n (f). The set of periodic points of the doubling map can be studied directly by solving the periodic point equation. Exercise.3.5. Find all points in P er n (f) for the doubling map f by directly solving the equation f n (x) = x. Let d be a distance on X, for example the arc length distance on S or d(x, y) = x y on the unit interval (see more about distances in the next chapter). n= Definition.3.3. We say that the set of periodic points of f : X X ϵ > 0 there exists x P er(f) such that d(x, y) < ϵ. is dense if for any y X and any Thus, if periodic points are dense, there exists periodic points arbitrarily close to any point of the space. Let us remark that density of the set of periodic points is not related to density of single orbits: if x is a periodic point, its orbit O + f (x) contains only finitely many points and thus cannot be dense if the space is infinite. By using the explicit list of periodic points (see Exercise above), one can show that periodic points for the doubling map are dense. c University of Bristol 00 & 04

13 Exercise.3.6. Prove that the periodic points for the doubling map are dense. [Hint: use the previous exercise.] There is another method to find periodic points for the doubling map, which uses the correspondence given by ψ and the knowledge of periodic points for the shift map, which are easy to describe. We illustrate now this method becouse it is very general and it allows to construct periodic points for maps for which the periodic point equation cannot be solved explicitely, but it is nevertheless possible to construct conjugacies with the shift by using coding. Periodic points for σ + are simply points (a i ) i= whose digits are repeated periodically. For example, if we repeat the digits 0,, periodically, we get a periodic point of period 3: (a i ) i= = 0,,, 0,,, 0,,, 0,,,... σ + ((a i ) i=) =,, 0,,, 0,,, 0,,,... (σ + ) ((a i ) i=) =, 0,,, 0,,, 0,,,... (σ + ) 3 ((a i ) i=) = 0,,, 0,,, 0,,,... = (a i ) i= More in general, if a n+i = a i for all i N, then so (a i ) i= is periodic of period n. (σ + ) n ((a i ) i=) = (a n+i ) i= = (a i ) i=, Theorem.3.. The doubling map f : R/Z R/Z has n periodic points of period n. Moreover, periodic points are dense. The second part of the theorem states that there are periodic points arbitrarily close to any other point. More precisely, for any x R/Z and any ϵ > 0 there exists a periodic point y such that d(x, y) < ϵ, where d is the distance on R/Z. Proof. Since the shift map σ + and the doubling map f are semi-conjugated, periodic points of period n for σ + : Σ + Σ + are mapped to periodic points of period n for f (not necessarily distint). Periodic points of period n for σ + are all sequences whose digits repeat periodically with period n. Thus, there are n such sequences, since we can choose freely the first n digits in {0, } and then repeat them periodically. Note that the sequence 0, 0, 0,... is mapped to 0 and the sequence,,,... is mapped to, which are the same point in R/Z = I/. Moreover, one can check that these are the only two periodic sequences which, applying Ψ, give identical periodic points 6. Thus, there are n periodic points of period n. The second part was proved in Exercise.3.6. Example.3.. The periodic points of period 3 for σ + are the periodic sequences obtained repeating the blocks of digits: , 0..., 0..., Let us take the corresponding binary expansions. Since a i+3 = a i for all i N we have a ( i i = a + a + a ) ( 3 a a 5 + a ) ( 3 a 6 + = + a + a ) 3 3 ( 3 ) j i= So, for example, starting from the sequence obtained repeating the block 0 we obtain ( ) 8 8 j = j = 5 ( ) 8 = j=0 Thus, we find the 7 = 3 periodic points of period 3 for the doubling map are 0, j=0 7, 7, 3 7, 4 7, 5 7, To see that, one can use that two binary expansions that produce the same number have one an infinite tail of digits 0 and the other an infinite tail of digits and the only periodic sequences which are eventually equal to 0 or eventually equal to are the sequence where all digits are 0 and the sequence where all digits are. j=0 c University of Bristol 00 & 04

14 .3.3 Symbolic coding for the doubling map Consider the two intervals P 0 = [ 0, ) [ ), P =,. They give a partition 7 {P 0, P } of [0, ]/, since P 0 P = and P 0 P = [0, ]/. Let ϕ : I/ Σ + be the map { x ϕ(x) = (a k ) ak = 0 if f k=0, where k (x) P 0, a k = iff k (x) P. The sequence a 0, a, a,..., a k,... is called the itinerary of O + f (x) (or, for short, the itinerary of x) with respect to the partition {P 0, P }: it is obtained by iterating f k (x) and recording which interval, whether P 0 or P, is visited at each k. In particular, if a 0, a, a,..., a k,... is called the itineary of O + f (x) we have x P a0, f(x) P a, f (x) P a,... f k (x) P ak,... Remark.3.. The idea of coding an orbit by recording its itineary with respect to a partition is a very powerful technique in dynamical systems. It often allow to conjugate a dynamical system to a shift map on a space of symbols. These symbolic spaces will be studied in Chapter and, even if at first they may seem more abstract, they are well studied and often easier to understand then the original system. Itineraries of the doubling map produce the digits binary expansions, in the following sense: Proposition. If a 0, a,..., a n,... is the itinerary of the point x [0, ], one has x = a 0 + a + a 3 + = that is a 0, a,..., a n,... are the digits of a 8 binary expansion of x. i= a i i, (.7) Before proving the proposition let us remark that another equivalent way to express that the doubling map gives the digits of a binary expansion of a point is the following: Corollary.3.. The map ϕ : R/Z Σ + is a right inverse for the map ψ : Σ + I/ constructed before, i.e. ψ ϕ : R/Z R/Z is the identity map. Indeed, given a point x [0, ), first ϕ(x) = (a i ) i=0 gives its itinerary a 0, a, a,... and then ψ((a i ) i=0 ) produces a point whose binary expansion has digits a i, which, by (.7) yield back the same point x. [The map ϕ is not a left inverse, i.e. the map ϕ ψ : Σ + Σ + is not necessarily the identity map. For most points, it is indeed identity, but the problem is related to points in X which have two binary expansions: the coding map ϕ yields only one of the two expansions.] Exercise.3.7. Give an example a point (a i ) i=0 such that ϕ(ψ((a i) i=0 ) (a i) i=0. [This shows that ϕ is not a right inverse for ψ. ] Proof of Proposition. Let a 0, a,..., a n,... be the itinerary of O + f (x) with respect to x. We have to check that it gives a binary expansion for x, that is that we can write x = a 0 + a + a 3 + = i= a i i. Step (i = 0) If the first digit of the itinerary is a 0 = 0, x P 0, that is 0 x < /. Thus the first digit of the binary expansion of x is 0 (since, if it were, we would have x = +... which is larger than /). Similarly, if a 0 =, x P, that is / x < and we can take as first digit of the binary expansion of x. 7 A finite (countable) partition of a set X is a subdivision of X into finitely many (respectively countably many) subsets X i X that are pairwise disjoint, that is X i X j = for all i j, and cover X, that is such that i X i = X. 8 Note that if x has two binary expansions, the itinerary will produce only one of them. 3 c University of Bristol 00 & 04

15 Step (i = k) To show that the k th entry a k gives the k th digit of the binary expansion of x, we can apply the doubling map k times and recall that if x, x,..., x k,... are digits of a binary expansion of x, since the doubling map acts as a shift on binary expansions, the digits of f k (x) are x k+, x k+,.... Moreover, by definition of itinerary, the itinerary of f k (x) is a k, a k+, a k+,.... Now, we can reason as before: if a k = 0 (respectively a k = ), then f k (x) P 0 (respectively P ) and 0 f k (x) < / (respectively / f k (x) < ). Thus, we see that the first digit of the binary expansion of f k (x), that is x k+, is 0 (respectively, can be taken to be ). Let a 0, a,..., a n be an (n + ) tuple of digit 0 or. Let I(a 0, a,..., a n ) = {x [0, ) such that ϕ(x) = (a 0, a,..., a n,... )} = {x [0, ) such that f k (x) P ak for all 0 k n} These are all points whose itinerary (and also whose binary expansion) starts with a, a,..., a n. In order to construct them, one can use that I(a 0, a,..., a n ) = P a0 f (P a ) f n (P an ). If x belongs to the intersection in the right hand side, clearly x P a0, f(x) P a,..., f n (x) P an, so by definition the itinerary starts with the block a 0, a,... a n. For example we have I(0) = P 0, I() = P and (see Figure.3) f (P ) = [ 4, ) [ 3 4, Repeating for the other pairs of digits we find: I() / ), so that I(0, ) = P 0 f (P ) = [ 4, ). 0 /4 / 3/4 I(0,) I(,) I(0, 0) = Figure.3: The intervals I(0, ) and I(, ) for the doubling map. [ 0, ) [, I(0, ) = 4 4, ) [, I(, 0) =, 3 ) [ ) 3, I(, ) = 4 4,. More in general, one can prove by induction that (P) each I(a 0, a,..., a n ) is an interval of length / n+ ; (P) as a 0, a,..., a n ranges through all possible (n + ) tuples of digit 0 or (which are n+, as many as choices of n + digits in {0, }), one obtains a partition of [0, ) into n+ intervals of length / n+ (as in the example for n = above): (a 0,...,a n ) {0,} n+ I(a 0, a,..., a n ) = [0, ). 4 c University of Bristol 00 & 04

16 Each interval is a dyadic interval of the form [ k n+, k + ) n+, where 0 k < n+. Let us use conjugacy and coding to construct a dense orbit for the doubling map. Theorem.3.. Let f be the doubling map. There exists a point x whose forward orbit O + f (x) under the doubling map is dense. Proof. Step We claim that to prove that an orbit O + f (x) is dense, it is enough to show that for each n it visits all intervals of the form I(a 0, a,..., a n ). Indeed, if this is the case, given y I and ϵ > 0, take N large enough so that / N+ ϵ and take the interval I(a 0, a,..., a N ) which contains y (one of them does since they partition [0, ) by (P) above). If we showed that there is a point f k (x) in the orbit O + f (x), which visits I(a 0, a,..., a N ), since both y and f k (x) belong to I(a,..., a N ) (which has size / N+ by (P) above), we have d(f k (x), y) / N+ < ϵ. This shows that O + f (x) is dense. Step. To construct an orbit which visits all dyadic intervals, we are going to construct its itinerary as a sequence in Σ + first. Let us now list for each n all the possible sequences a 0, a,..., a n of length n (there are n+ of them) and create a sequence (a i ) i=0 by just apposing all such sequences for n = 0, then n =, then n = and so on: 0,, 0, 0, 0,,, 0,, 0,,, 0, 0, 0, 0, 0,, 0,, 0, 0,,,, 0, 0,... Step 3. Let us consider now the point x := ψ((a i ) i=0 ) which has the sequence (a i) i=0 as digist of its binary expansion. Let us show that its orbit visits all intervals of the form I(a 0, a,..., a n ) and thus that it is dense by Step. To see that, it is enough to find where the block a 0, a,..., a n appears inside (a i ) i=0, for example at a k = a 0, a k+ = a,..., a k+n = a n. Then, since the itinerary of f k (x) by definition of itinerary is a k, a k+,..., a k+n,..., this shows that f k (x) I(a k, a k+,..., a k+n ) = I(a 0, a,..., a n ), so f k (x) is the point in O + f (x) which visits I(a 0, a,..., a n ). This concludes the proof that O + f (x) visits all dyadic intervals and hence that it is dense. Exercise.3.8. Draw all intervals of the form I(a, a, a 3 ) where a, a, a 3 {0, }. Linear expanding maps. We remarked earlier that the doubling map doubles distances: if x, y R/Z are any two points such that d(x, y) < /4, then d(f(x), f(y)) = d(x, y), that is, the distance of their images is doubled. The doubling map is an example of an expanding map: Definition.3.4. A one-dimensional map g : I I of an interval I R is called an expanding if it is piecewise differentiable, that is we can decompose I into a finite union of intervals on each of which g is differentiable, and the derivative g satisfies g (x) > for all x I. More precisely, the doubling belongs to the family of linear expanding maps of the circle: for each m Z with m > the map E m : R/Z R/Z is given by E m (x) = mx mod (or equivalently E m (z) = z k on S ). The doubling map is the same than E. These maps can be studied analogously, by considering expansion in base m instead than binary expansions. One can prove that they are semi-conjugated with the shift σ + on the space Σ + m = {0,,..., m } N of one-sided sequences in the digits 0,..., m. 5 c University of Bristol 00 & 04

17 .4 Baker s map Let [0, ) = [0, ) [0, ) be the unit square. Consider the following two dimensional map F : [0, ) [0, ) { ( ) x, y F (x, y) = ( ) x, y+ if 0 x <, if x <. Geometrically, F is obtained by cutting [0, ) into two vertical rectangles R 0 = [0, /) [0, ) and R = [/, ) [0, ), stretching and compressing each to obtain an interval of horizontal width and vertical height / and then putting them on top of each other. The name baker s map comes becouse this mimic the F / F 0 / 0 Figure.4: The action of the baker s map. movement made by a baker to prepare the bread dough 9. Similar maps are often use in industrial processes since, as we will see formally later, they are very effective in quickly mixing. Remark.4.. Notice while the horizontal direction is stretched by by a factor, the vertical direction is contracted by a factor /. The baker map is invertible. The inverse of the map F can be explicitely given by { ( x F (x, y) = (, y) if 0 y <, x+, y ) if y <. Geometrically, F cuts X into two horizontal squares and stretches each of them to double the height and divide by two the width and then places them one next to each other (in Figure.4, the right square now gives the departing rectangle decomposition and the left square shows the images of each rectangle under F ). Unlike in the case of the doubling map, we now have to be more careful in identifying F as a map on X = (R/Z). The map F : X X ( ) {x}, {y} if 0 {x} <, F (x, y) = ( {x}, {y}+ ) if {x} <. is well defined; here {y} = y mod denotes the fractional part of y. If you compare the definition of the baker map F with the doubling map f in the previous section, you will notice that the horizontal coordinate is trasformed exactly as f. More precisely one can show that F is an extenstion of f (i.e. there is a semiconjugacy ψ such that ψ F = f ψ, see Exercise below). Extensions of non-invertible maps which are invertible are called intervible extensions or natural extensions. Exercise.4.. Show that the doubling map f : R/Z R/Z and the baker map F : (R/Z) (R/Z) are semi-conjugated and the semi-conjugacy is given by the projection π : (R/Z) R/Z given by π(x, y) = x. 9 There are other versions of the baker s map where the dough is not cut, but folded over. 6 c University of Bristol 00 & 04

18 To study the doubling map, we introduced the one-sided shift on two symbols (σ + : Σ + Σ + ). To study the baker map is natural to introduce the bi-sided shift on two symbols, that we now define. Let Σ = {0, } Z be the set of all bi-infinite sequences of 0 and : Σ = {(a i ) i=, a i {0, }}. A point a Σ + is a bi-sided sequence of digits 0,, for example... 0, 0,,, 0,, 0, 0, 0,,, 0,,,,, 0, 0,, 0,, 0... The (bi-sided) shift map σ is a map σ : Σ Σ which maps a sequence to the shifted sequence: σ((a i ) i= ) = (b i ) i=, where b i = a i+. (.8) The sequence (b i ) i= is obtained from the sequence (a i) i= by shifting all the digits one place to the left. For example, if (a i ) i= (b i ) i= =... 0, 0,,, 0,, 0, 0, 0,,, 0,,,,,... = 0, 0,,, 0,, 0, 0, 0,,, 0,,,,,... Note that while the one-sided shift σ + was not invertible, because we were throwing away the first digit a of the one-sided sequence (a i ) i= before shifting to the left, the map σ is now invertible. The inverse σ is simply the shift to the right. One can show that the baker map F and the bi-sided shift σ are semi-conjugate if σ is restricted to a certain shift-invariant subspace (see Theorem.4. below). In the case of the doubling map, the key was to use binary expansion. What to use now? We can get a hint of what is the semi-conjugacy using itineraries and trying to understand sets which share a common part of their itinerary. Let R 0 and R be the two basic rectangles [ R 0 = 0, ) [0, ), R = [ ), [0, ). (See Figure.4, left square: R 0 is the left rectangle, R the right one.) The (bi-infinite) itinerary of (x, y) with respect to the partition {R 0, R } is the sequence (a i ) + i= Σ given by { ak = 0 if F k (x, y) R 0, a k = if F k (x, y) R In particular, if..., a, a, a 0, a, a,... is the itineary of O F ((x, y)) we have F k (x, y) R ak, for all k Z. Note that here, since F is invertible, we can record not only the future but also the past. Let us now define sets of points which share the same finite piece of itinerary. Given n, m N and a k {0, } for m k n, let R m,n (a m,..., a n ) = {(x, y) X F k (x, y) R ak for m k n}. These are all points such that the block of the itineary from m to n is given by the digits a m, a m+,..., a, a 0, a,..., a n, a n. To construct such sets, let us rewrite them as n R m,n (a m,..., a n ) = F k (R ak ). k= m 7 c University of Bristol 00 & 04

19 Example.4.. Let us compute F (R 0 ). Either from the definition or from the geometric action of F, one can see that (see Figure.5(a)) [ F (R 0 ) = 0, ) [ [0, ) 4, 3 ) [0, ). 4 Thus, (see Figure.5(b)) R 0, (, 0) = R F (R 0 ) = [, 3 ) [0, ). 4 0 /4 / 3/4 (a) F (R 0 ) 0 /4 / 3/4 (b) R 0, (, 0) 0 /8 /4 3/8 / 5/8 3/4 7/8 (c) R 0, (a, a, a ) Figure.5: Examples of rectangles determined by future itineraries. One can prove that all rectangles deterimed by forward itineraries, i.e. of the form R 0,n (a 0, a,..., a n ), are thin vertical rectangles of width / n+ and full height, as in Figure.5(c), and as a 0,..., a n changes, they cover X. More precisely, recalling the intervals I(a 0,..., a n ) defined for the doubling map 0, we have [ k R 0,n (a 0, a,..., a n ) = I(a 0,..., a n ) [0, ) = n+, k + ) n+ [0, ) for some 0 k < n+. Let us now describe a set which share the same past itinerary. Example.4.. The image F (R 0 ) is the bottom horizontal rectangle in the left square in Figure.4. The image F (R 0 ) is shown in Figure.6(a) and is given by [ F (R ) = [0, ) 0, ) [ [0, ) 4, 3 ). 4 [Try to convince yourself by imagining the geometric action of F on these sets (or by writing an explicit formula)]. Hence, for example (see Figure.6(b)) [ R, (0, ) = F (R 0 ) F (R ) = [0, ), 3 ). 4 In general, one can verify that each set of the form R n, (a n,..., a ) (dependend only on the past itinerary) is a thin horizontal rectangle, of height / n and full width, as in Figure.6(c). Exercise.4.. Draw the following sets: (a) R,0 (0, ) 0 This is because the future history of F, i.e. whether F k (x, y) with k 0 belongs to R 0 or R, is completely deterimed by the doubling map. 8 c University of Bristol 00 & 04

20 3/4 / 3/4 / 7/8 3/4 5/8 / /4 0 0 (a) F (R 0 ) /4 0 0 (b) R, (0, ) 3/8 /4 /8 0 0 (c) R 3, (a 3, a, a ) Figure.6: Examples of rectangles determined by past itineraries. (b) R, (0,, ) (c) R,0 (, 0, ) In general R m,n (a m,..., a n ) is a rectangle of horizontal width / n+ and height / m. The more we precise the backwards itinerary a, a,..., a n, the thinner the precision with which we determine the vertical component y. Moreover, from the geometric picture, you can guess that as a 0, a,..., a n,... give the digits of the binary expansion of x, a, a,..., a n,... give the digits of the binary expansion of y. This is exactly the insight that we need to construct the semi-conjugacy with the full shift. Now we are ready to construct a semi-conjugacty beteween the baker map and the full shift which is a conjugacy outside a measure zero set of points. Denote by T = {a Σ : i 0 Z such that a i = i > i 0 } the set of sequences with forward tails consisting only of s. We have σ(t ) = T, i.e., T is a shift-invariant subspace. Hence its complement Σ = Σ \ T is also shift-invariant. Theorem.4.. The baker map is semi-conjugated to the full shift σ : Σ Σ via the map Ψ : Σ X given by Ψ((a i ) + i= ) = (x, y) where x = i= a i i mod, y = i= a i i mod. As for the doubling map, binary expansions turns out to be crucial to build the map Ψ. While the future (a i ) i=0 of the sequence (a i) i= will be used to give the binary expansion of x, the past (a i) i= of the sequence (a i ) i= turns out to be related to the binary expansion of the vertical coordinate y. Proof of Theorem.4.. For every point (x, y), both x and y can be expressed in binary expansion. If x has a binary expansion of the form a 0,..., a i0, 0,,,,,,... (i.e., with a forbidden tail), then x also has the binary expansion a 0,..., a i0,, 0, 0, 0, 0, 0,... (exercise!). This shows that Ψ is surjective. Thus it remains to check that Ψσ = F Ψ. Let us first compute Ψ ( σ((a i ) + i= )) = Ψ ( ( (a i+ ) + ) a i i= = i mod, i= i= a i+ i mod Let a 0 be the first digit of the binary expansion of x. We have that a 0 = 0 if 0 x < and a 0 = if x <. For x = we have either a 0 = (then a i = 0 for i ) or a 0 = 0 (then a i = for i ). The latter has a forbidden tail and thus does not occur. Therefore a 0 = 0 if 0 x < and a 0 = if x <, which allows us to write ( F (x, y) = x mod, y + a ) 0. ). 9 c University of Bristol 00 & 04

21 We compute F (Ψ((a i ) + i= )) = F ( = ( which is the same answer as for Ψσ. Extra: Coding in dynamics i= i= a i i, i= a i i ) a i i mod, a 0 + i= a i i ) = ( i= a i i, i= a i+ i The idea of coding is very powerful in dynamical systems. Let f : X X be a dynamical system and X a partition of X in finite (or sometimes countably) many pieces, that is a collection of sets {P 0,..., P N }, P i P j = i j, and N i=0 P i = X. Then we can defined the itinerary of x X with respect to the partition {P 0,..., P N } as a sequence of digits a k such that f k (x) P ak, for all k. Here the digits a k {0,..., N} (they take as many values as the number of elements in the partition). If f is not invertible, it makes sense to consider forward itineraries, so we get a sequence (a k ) k= {0,..., N}N. If f is invertible, we can also consider backwards itineraries, so we get a sequence (a k ) + k= {0,..., N}Z. If x has itinerary (a k ) k= {0,..., N}N (or respectively (a k ) + ) then the itinerary of f(x) will be shifted, so it will be given by σ + ((a k ) k= ) (or respectively σ ( (a k ) + k= ) ). Thus, the coding maps an orbit of f into an orbit of the shift on a symbolic space. Two important questions to ask are: Does the itinerary completely determine the point x? Are all itineraries possible, i.e. does any sequence in Σ correspond to an actual itinerary? In the case of the doubling map and of the baker map, both questions had positive answer. This allowed us to get a semi-conjugacy with the shift (one-sided or bi-sided). The itinerary is often enough to completely determine the point x: this is the case when the maps we are looking at are expanding. On the other hand, many times not all sequences arise as itineraries: we will describe in Chapter 3 more general symbolic spaces (subshifts of finite type) which will capture the behavior of many more maps..5 Hyperbolic toral automorphisms Consider the unit square [0, ] [0, ]. If you glue the two parallel vertical sides and the two parallel horizontal sides by using the identifications (x, 0) (x, ), x [0, ], (0, y) (, y), y [0, ], (.9) we get the surface of a doughnut, which is called a two-dimensional torus and denoted by T (when you glue the vertical sides first, you get a cylinder, then when you glue the horizontal sides you are gluing the two circles which bound the cylinder and you get the torus). More formally, T = R /Z = R/Z R/Z is the set of equivalence classes of points (x, y) R modulo Z : two points (x, y) and (x, y ) in R are equivalent and we write (x, y) (x, y ) if and only if (x x, y y ) = (k, l), where (k, l) Z. The unit square contains at least one representative for each equivalence class. Note though that since the points (x, 0) and (x, ) for x [0, ] differ by the vector (0, ) they represent the same equivalence class. ), 0 c University of Bristol 00 & 04

22 Similarly, the points (0, y) and (, y) for y [0, ], which differs by the vector (, 0), also represent the same equivalence class, while all interior points represent distinct equivalence classes. Thus, the square with opposite sides glued by the identifications (.9) contains exactly one representative for each equivalence class. In this sense, the space R /Z of equivalence classes is represented by a square with opposite sides identified as in (.9) We will now study a class of maps of the torus. Let A be a matrix with integer entries. Then A acts linearly on R (and in particular on the unit square): ( ) ( ) ( ) a b x ax + by if A =, where a, b, c, d Z, then A =. c d y cx + dy We claim that we can use this linear action to define a map f A : T T on the torus T = R /Z, given by f A (x, y) = (ax + by mod, cx + dy mod ). (.0) We have to check that if (x, y) (x, y ), then f A (x, y) f A (x, y ). But (x, y) (x, y ) means that (x x, y y ) = (k, l) where (k, l) Z and since A has integer entries, it maps integer vectors to integer vectors so ( ) ( ) ( ) x x k A A y y = A Z. l Thus, since f A is defined taking entries mod, we have f A (x, y) = f A (x, y ) as desired. The map f A : T T is not invertible in general. However, if det(a) = ad bc = ±, then ( ) ( ) A d b d b = = ± det(a) c a c a is again a matrix with integer entries, so it induces a well-defined map f A : T T. Thus, if det(a) = ±, then f A is invertible and the inverse is given by (f A ) = f A. Definition.5.. If A is an integer matrix, we say that f A : T T defined in (.0) is a toral endomorphism. If furthermore det(a) = ±, we say that f A : T T is a toral automorphism. Example.5. (Arnold CAT map). Let A be the matrix ( ). Then the induced map f A : T T is f A (x, y) = (x + y mod, x + y mod ). This map is known as Arnold s CAT map. To draw the action of the map, consider first the image of the unit square by the linear action of A: since the two bases vectors e, e are mapped to ( ) ( ) ( ) ( ) ( ) ( ) 0 A e = =, A e 0 = =, the image of the unit square is the parallelogram generated by Ae, Ae (see Figure.7). To visualize f A we need to consider the result modulo in each coordinate (or, equivalently, modulo Z ), which means subtract to each point an integer vector (the vector whose components are the integer-parts of the x and y coordinate respectively) to get an equivalent point in unit square. Geometrically, this means cutting and pasting different triangles (each one consisting of all points of the parallelogram contained in a square of the unit square grid) back to the standard unit square, as shown in Figure.7. In many books CAT is used as a shortening for Continuous Automorphism of the Torus. The first time the map was described in the book by Arnold, though, the action of the map is illustrated by drawing the face of a cat, as in Figure.7. c University of Bristol 00 & 04

23 Figure.7: The Arnold CAT map. Definition.5.. The toral automorphism f A : T T given by a matrix A with integer entries and determinant ± is called hyperbolic if A has no eigenvalues of modulus. Remark.5.. Note that if det(a) =, if λ > is one of the eigenvalues of A, the other eigenvalues is λ = /λ (since the product of the eigenvalues is the determinant). Thus, if one eigenvalue is greater than in modulus, the other is automatically less than in modulus. Example.5.. The matrix A has eigenvalues λ = > and λ = = 3 5 <. Thus, f A is a hyperbolic toral automorphism. The corresponding eigenvectors v and v are ( + 5 ) ( 5 v =, v = Thus, A expands all lines in the direction of v by λ > and contracts all lines in the direction of v by multiplying them by λ <. Since f A is obtained by cutting and pasting the image of the unit square by A, the same is true for f A (see Figure.7). Let us study periodic points of hyperbolic toral automorphisms. Theorem.5.. Let f A : T T be a hyperbolic toral automorphism. The periodic points of f A are exactly all points in [0, ) [0, ) which have rational coordinates, that is all points ( p q, p ), p, p N, q N\{0}, 0 p, p < q. q ). c University of Bristol 00 & 04

24 ( ) Proof. If (x, y) = p q, p q has rational coordinates, then f n A(x, y) = ( p (n) q ), p(n), q where p (n), p(n) are integers 0 p (n), p(n) < q which are given by p (n) q p (n) q = A n ( p q p q ) mod Z = ( anp +b np q mod c np +d np q mod ), if A n = ( ) an b n, c n d n that is, which are given by ( p (n) p (n) ) ( an p = + b n p mod q c n p + d n p mod q ) ( ), where A n an b = n. c n d n Since there are at most q choices for the pair (p (n), p(n) ), there exists 0 m n q + such that fa m(x, y) = f A n(x, y). This means that (x, y) is eventually periodic and, since f A is invertible, this implies that (x, y) is periodic (see Exercise.(b)). Conversely, if (x, y) is periodic, there is n such that fa n(x, y) = (x, y) and by definition of f A this means that there exists k, l Z such that ( ) ( ) ( ) ( ) ( ) ( ) A n x x k = + (A n x k 0 I) =, where I =. y y l y l 0 Let us check that (A n I) is invertible. Since f A is hyperbolic, A has no eigenvalues of modulus, so A n has no eigenvalues. Thus, there is no non-zero vector v that solves (A n I)v = 0 and this shows that (A n I) is invertible. Hence, we can solve: ( x y ) ( = (A n I) k l Since A has entries in Z, (A n I) has entries in Z and (A n I) has entries in Q, so both x, y are rational numbers. See the Extra for an application of the previous thereom to explain why discretization of the CAT map (or in general, of a toral automorphism), for example by digital pixels on a screen, will eventually appear to be periodic. One can precisely compute the number of periodic points of period n in a hyperbolic toral automorphism. Theorem.5.. Let f A : T T be a hyperbolic toral automorphism associated to a matrix A with det(a) = and eigenvalues λ and λ. The number of periodic points of period n is λ n + λ n. Proof. Fixed points of period n are solutions of fa n (x, y) = (x, y). Equivalently, as we just proved, ) ( (A n x I) y = ( k l ). ), for some k, l Z. (.) The map (A n I) maps the unit square [0, ) [0, ) to a parallelogram P generated by the vectors (A n I)e and (A n I)e. Thus, solutions to (.) correspond to integer points (k, l) Z which belong to the parallelogram P. The number of such integer points is exactly the area of P in virtue of the following result known as Pick s theorem: Theorem.5.3 (Pick s Theorem). Consider a parallelogram P R whose vertices are integer points in Z and let P/ be obtained by identifying opposite parallel sides of P. Then the number of points in P/ which integer coordinates is exactly equal to the area of P. 3 c University of Bristol 00 & 04

25 A more general form of Pick s theorem and its proof are included as an Extra. Thus, the number of integer points in P is exactly the area of P which is given by det(a n I). Since the determinant is the product of the eigenvalues, it is enough to compute the eigenvalues of (A n I). If v is an eigenvector of (A n I) with eigenvalue µ, then (A n I)v = µv A n v = (µ + )v, so v is an eigenvector for A n with eigenvector µ +. Since A has eigenvectors λ, λ, A n has eigenvectors λ n, λ n and then (A n I) has eigenvectors λ n, λ n. So, putting everything together we proved that Card{(x, y) s.t. f n A(x, y) = (x, y)} = Area(P ) = det(a n I) = (λ n )(λ n ) = λ n + λ n, where in the last inequality we used that λ n λ n =, since λ n and λ n are the eigenvalues of A n, which has determinant since det(a n ) = det(a) n and det(a) = by assumption. Remark.5.. All definitions in this section generalize to higher dimensions. The k-dimensional torus T k is R k /Z k. Given a k k matrix A with integer entries, it determines a map f A : T k T k. If det(a) =, f A is called a toral automorphism. If A has no eigenvalues of modulus, f A is an hyperbolic toral automorphism. Periodic points for f A are exactly points with rational coordinates..6 Gauss map and continued fractions In this lecture we will introduce the Gauss map, which is very important for its connection with continued fractions in number theory. The Gauss map G : [0, ] [0, ] is the following map: { 0 if x = 0 G(x) = { } x = x mod if 0 < x Here {x} denotes the fractional part of x. We can write {x} = x [x] where [x] is the integer part. Equivalently, {x} = x mod. Note that [ ] = n n x x < n + n + < x n. Thus, explicitely, one has the following expression (see the graph in Figure.8): { 0 if x = 0 G(x) = x n if n+ < x for n N. n The rescrition of G to an interval of the form (/n +, /n] is called branch. Each branch G : (/n +, /n] [0, ) is monotone, surjective (onto [0, )) and invertible (see Figure.8). The Gauss map is important for its connections with continued fractions. A finite continued fraction (CF will be used as shortening for Continued Fraction) is an expression of the form (.) a 0 + a + a + a an where a 0, a, a,..., a n N\{0} are called entries of the continued fraction expansion. We will denote the finite continued fraction expansion by [a 0, a, a,..., a n ]. Every finite continued fraction expansion correspond to a rational number p/q (which can be obtained by clearing out denominators). Example.6.. For example + 3 = 3+ 3 = c University of Bristol 00 & 04

26 /4/3 / Figure.8: The first branches of the graph of the Gauss map. Conversely, all rational numbers in [0, ] admit a representation as a finite continued fraction. Example.6.. For example 3 4 = +, = Every irrational number x (0, ) can be expressed through a (unique) infinite continued fraction 3, that we denote by [a 0, a, a, a 3,... ] =. a 0 + a + a + a Example.6.3. For example π = 3 +, = The number ( 5 )/ is known as golden mean 4 and it has the lowest possible continued fraction entries, all entries equal to one. Similarly, the number whose CF entries are all equal to is known as silver mean. This representatin is not unique: if the last digit a n of a finite CF is, then [a 0,..., a n, ] = [a 0,..., a n + ]. If one requires that the last entry is different from one, though, then one can prove that the representation as finite continued fraction is unique. 3 To be precise, when we write such an infinite continued fraction expression, its value is the limit of the finite continued fraction expansion truncations [a 0, a, a, a 3,..., a n ], each of which is a well defined rational number. One should first prove that this limit exist, see (.3). 4 The inverse of the golden mean is 5+, known as golden ratio. It appears often in art and in nature since it is considered aesthetically pleasing: for example, the ratio of the width and height of the facade of the Partenon in Athens is exactly the golden ratio and a whole Renessaince treaty, Luca Pacioli s De divina proportione, written in 509, is dedicated to the golden ratio in arts, science and architecture. 5 c University of Bristol 00 & 04

27 One can see that a number is rational if and only if the continued fraction expansion is finite. If x is an irrational number whose infinite continued fraction expansion is [a 0, a, a,... ], one can truncate the continued fraction expansion at level n and obtain a rational number that we denote p n /q n p n q n = [a 0, a, a,..., a n ]. These numbers p n /q n are called convergents of the continued fraction. Two of the important properties of convergents are the following:. One can prove that p n /q n converge to x exponentially fast, i.e. p n lim = x and p n n q n x q n ( ). (.3) n Thus, the fractions p n /q n give rational approximations of x.. Convergents give best approximations among all rational approximations with denominator up to q n, that is x p n x p q, p Z, 0 q q n. q n One can also see that the continued fraction expansion of an irrational number is unique. To find the continued fraction expansion of a number, we will exploit the relation with the symbolic coding of the Gauss map, in the same way that binary expansions are related to the symbolic coding of the doubling map. Let P n be the subintervals of [0, ) naturally determined by the domains of the branches of the Gauss map: ( ] P =,, P = ( 3, ], P 3 = ( 4, 3 ],..., P n = ( n +, n ],... Note that P n accumulate towards 0 as n increases If we add P 0 = {0}, the collection {P 0, P,..., P n,... } is a (countable) partition 5 of [0, ]. Theorem.6.. Let x be irrational. partition {P 0, P, P,..., P n,... }, i.e. Let a 0, a,..., a n,... be the itinerary of O + G (x) with respect to the x P a0, G(x) P a,..., G (x) P a,..., G k (x) P ak,..., Then x = [a 0, a, a,..., a n,... ]. Thus, itineraries of the Guass map give the entries of the continued fraction expansions. Remark.6.. If x is rational, then there exists n such that G n (x) = 0 and hence G m (x) = 0 for all m n. In this case, G m (x) P 0 for all m n so the itinerary is eventually zero. The theorem is still true if we consider the beginning of the itineary: the finite itinerary before the tail of 0 gives the entries of the finite continued fraction expansion of x. Proof. Let us first remark that x P n n + < x n n [ ] x < n + = n. (.4) x In particular, a 0 = [/x] since x P a0. Thus, { } G(x) = = [ ] x x = x x a 0 x = a 0 + G(x). 5 Recall that a partition is a collection of disjoint sets whose union is the whole space. 6 c University of Bristol 00 & 04

28 Let us prove by induction that [ ] a n = G n and x = (x) a 0 + a +... an+g n+ (x) = [a 0, a,..., a n + G n+ (x)]. (.5) We have already shown that this is true for n = 0. Assume [ that it is] proved for n and consider n +. Since G n+ (x) P an+ by definition of itinerary, we have a n+ = by (.4). This proves the first part of (.5) for n +. Then, recalling the definition of G we have [ ] G n+ (x) = G n+ (x) G n+ = (x) G n+ (x) G n+ (x) a n+ G n+ (x) = so that, plugging that in the second part of the inductive assumption (.5) we get x = a a n+g n+ (x) =, a a n+ a n+ +G n+ (x) a n+ + G n+ (x) which proves the second part of (.5) for n +. Thus, recursively, the itinerary is producing 6 the infinite continued fraction expansion of x. From the proof of the previous theorem, one can see the following. Remark.6.. The Gauss map acts on the digits of the CF expansion as the one-sided shift, that is if x = [a 0, a, a,..., a n,... ] then G(x) = [a, a, a 3,..., a n+,... ]. One can characterize in terms of orbits of the Gauss map various class of numbers. For example:. Rational numbers are exactly the numbers x which have finite continued fraction expansion or equivalently such that there exists n N such that G n (x) = 0 (eventually mapped to zero by the Gauss map).. Quadratic irrationals, that is numbers of the form a+b c d, where a, b, c, d are integers 7, are exactly numbers which have a eventually periodic continued fraction expansion or equivalently are pre-periodic points for the Gauss map. In number theory (and in particular in Diophantine approximation) other class of numbers (for example Badly approximable numbers) can be characterized in terms of their continued fraction expansion 8. Example.6.4. We have already seen two examples of quadratic irrationals, the golden mean g and the silver mean s: 5 g = =, s = + = Both the golden mean and the silver mean are fixed points of the Gauss map: G(g) = g, G(s) = s. Similarly all other fixed points correspond to numbers whose continued fraction entries are all equal. 6 One should still prove that the finite continued fractions in (.5) do converge, as n tends to infinity and that the limit is x. This can be done by the same method that one can use to show that convergents tend to x exponentially fast. 7 Equivalently, one can define quadratic irrationals as solutions of equations of degree two with integer coefficients. 8 One can defined Badly approximable numbers as the numbers for which there exists a number A such that all entries a n of their continued fraction expansion are bounded by A. In particular, quadratic irrationals are badly approximable. 7 c University of Bristol 00 & 04

29 Example.6.5. Let α = Then one can check that α = [, 3,, 3,, 3,... ], so that the entries are periodic and the period is. Thus G (α) = α. Explicitely, since we know the itinerary of α, we can write down the equation satisfied by α. We know that G(α) = α, since [ α so that the equation G (α) = α becomes ] =, and G(G(α)) = G(α) 3 α 3 = α. since [ ] = 3, G(α) Using the ideas in the previous exercise, one can produce quadratic irrationals with any given periodic sequence of CF entries. Exercise.6.. Prove that if G n (x) = 0 then x has a representation as a finite continued fraction expansion and thus it is rational. * Exercise.6.. Prove that if G n (x) = x then x satisfies an equation of degree two with integer entries. Conclude that x is a quadratic irrational. 8 c University of Bristol 00 & 04

30 Appendix A Extras to Part I A. Extras to Sec.. A.. Continuous Dynamical Systems A continous dynamical system can be given by a parameter family of maps f t : X X where t R. The main example is given by solutions of a differential equation. Let X R n be a space, g : X R n a function, x 0 X an initial condition and { ẋ(t) = g(x) (A.) x(0) = x 0 be a differential equation. If the solution x(x 0, t) is well defined, unique and exists for all t and all initial conditions x 0 X, if we set f t (x 0 ) := x(x 0, t) we have an example of a continous dynamical system. In this case, an orbit is given by the trajectory described by the solution: Definition A... If {f t } t R is a continuous dynamical system, we denote O ft (x) := {f t (x), t R}. More in general, a -parameter family {f t } t R is called a flow if f 0 is the identity map and for all t, s R we have f t+s = f t f s, i.e. A.. f t+s (x) = f t (f s (x)) = f s (f t (x)), for all x X. Dynamical systems as actions A more formal way to define a dynamical system is the following, using the notion of action. Let X be a space and G group (as Z or R or R d ) or a semigroup (as N). Definition A... An action of G on X is a map ψ : G X X such that, if we write ψ(g, x) = ψ g (x) we have () If e id the identity element of G, ψ e : X X is the identity map; () For all g, g G we have ψ g ψ g = ψ g g. A discrete dynamical systems is then defined as an action of the group Z or of the semigroup N. A continous dynamical system is an action of R. There are more complicated dynamical systems defined for example by actions of other groups (for example R d ). If X has an additional structure (for example X is a topological space or X is a measured space), we can ask the additional requirement that for each g G, ψ g : X X preserves the structure of X (for example ψ g is a continous map if X is a topological space or ψ g preserves the measure. We will see more precisely these definitions in Chapters and 4. i

31 Exercise A... Prove that the iterates of a map f : X X give an action of N on X. N X X is given by (n, x) f n (x). Prove that if f is invertible, one has an action of Z. The action Exercise A... Prove that the solutions of a differential equation as (A.) (assuming that for all points x 0 X the solutions are unique and well defined for all times) give an action of R on X. A. Extra to Section. A.. Dirichlet s theorem Let us show an application to number theory. The following result is known as Dirichlet s theorem: Theorem A.. ((Dirichlet Theorem)). If α is irrational, for each δ > 0 there exists an integer q with 0 < q /δ and p Z such that α p q δ q The theorem shows how well an irrational number can be approximated by rational numbers. It is a first important keystone in the theory of Diophantine Approximation. Proof. Note that in additive notation the orbit of 0 contains the points R n α(0) = nα mod. We saw in the proof of Theorem.. that when α is irrational, the orbit O + R α (0) consists of distinct points. If we choose N large so that /N δ, by Pigeon Hole principle there exists 0 n < m N such that d(r m (0), R n (0)) < N. Let us call q = m n, so 0 < q N. Since R α is an isometry, d(0, R q (0)) = d(0, R m n (0)) = d(r n (0), R m n (R n (0))) = d(r m (0), R n (0)) < N. Recalling the definition of distance, this means exactly that there exists an integer p Z such that qα p N δ. As a Corollary, one can show the following version of Dirichlet theorem: Exercise A... Show that if α is irrational, there are infinitely many fractions p/q where p Z, q N that solve the equation α p q q. A.3 Extra to Section.3 A.3. A conjugacy for the doubling map. In order to obtain a conjugacy between the doubling map f : X X and a shift is to consider the shift space Σ/ with the equivalence relation a b if and only if there exists a k N such that a i = b i for 0 i < k and a k =, a i = for all i > k, b k = 0, b i = for all i > k. The sequences which are identified are sequences with tails of 0s and s which correspond to two possible choices of binary digits for a dyadic rational. ii c University of Bristol 00 & 04

32 Theorem A.3.. The map ψ : Σ/ X is well defined and it is a conjugacy between the doubling map f and σ : Σ/ Σ/. Exercise A.3.. Prove Theorem A.3.. Exercise A.3.. Let D be the set of dyadic rationals and let Σ be the sequences which do not end with a tail of 0s or s. Verify that ψ : Σ X D is a conjugacy. A.4 Extras to Section.5 A.4. A discrete toral automorphism As an illustration of Theorem on periodic points of hyperbolic toral automorphisms, one can see Figure A., in which iterates of a discretization of the Cat Map is plotted. If the image is made by a finite number of pixels, say q, each representing a point (p /q, p /q) with rational coordinates, then an iterate of f A is the identity by Theorem.5. proved in this section. Figure A.: A discrete version of the Arnold CAT map. A.4. Pick s theorem on the area of a parallelogram with integer vertices Let us state and prove Pick s Theorem (which was used in the computation of the number of periodic points for a hyperbolic toral automorphism). Theorem A.4. (Pick s Theorem). The area of a parallelogram P R whose vertices are integer points in Z is given by the number of points of Z which are contained, where points on the edges are counted as half and all vertices count as a single point. More precisely, let i be the number of points with integer coordinates in the interior of P and b be the number of points with integer coordinates on the perimeter of P. Then, Area(P ) = i + (b/) +. Let us remark that this form implies as a corollary the formulation used in the lecture notes. Indeed, consider P/ obtained by glueing opposite parallel sides of P. Then the points with integer coordinates on the sides of P are glued together in pairs, and all the vertices of P are identifyed to produce the equivalence class of a unique point on P/. Thus, the points with integer coordinates in P/ are exactly i + (b/) +. so the number of integer po P/ Before the proof, let us present some lemmas. Lemma A.4.. If a polygon P as in the assumptions of Pick s theorem is divided into two smaller polygons, P and P by a path whose endpoints also belong to Z, then, if Pick s formula holds for both P and P, it also holds for P. iii c University of Bristol 00 & 04